LEALLA is a collection of lightweight language-agnostic sentence embedding models supporting 109 languages, distilled from LaBSE. The model is useful for getting multilingual sentence embeddings and for bi-text retrieval.

Model: HuggingFace's model hub.
Paper: arXiv.
Original model: TensorFlow Hub.
Conversion from TensorFlow to PyTorch: GitHub.

This is migrated from the v1 model on the TF Hub. The embeddings produced by both the versions of the model are equivalent. Though, for some of the languages (like Japanese), the LEALLA models appear to require higher tolerances when comparing embeddings and similarities.

Usage

Using the model:

python

import torch
from transformers import BertModel, BertTokenizerFast


tokenizer = BertTokenizerFast.from_pretrained("setu4993/LEALLA-large")
model = BertModel.from_pretrained("setu4993/LEALLA-large")
model = model.eval()

english_sentences = [
    "dog",
    "Puppies are nice.",
    "I enjoy taking long walks along the beach with my dog.",
]
english_inputs = tokenizer(english_sentences, return_tensors="pt", padding=True)

with torch.no_grad():
    english_outputs = model(**english_inputs)

To get the sentence embeddings, use the pooler output:

english_embeddings = english_outputs.pooler_output

Output for other languages:

python

italian_sentences = [
    "cane",
    "I cuccioli sono carini.",
    "Mi piace fare lunghe passeggiate lungo la spiaggia con il mio cane.",
]
japanese_sentences = ["犬", "子犬はいいです", "私は犬と一緒にビーチを散歩するのが好きです"]
italian_inputs = tokenizer(italian_sentences, return_tensors="pt", padding=True)
japanese_inputs = tokenizer(japanese_sentences, return_tensors="pt", padding=True)

with torch.no_grad():
    italian_outputs = model(**italian_inputs)
    japanese_outputs = model(**japanese_inputs)

italian_embeddings = italian_outputs.pooler_output
japanese_embeddings = japanese_outputs.pooler_output

For similarity between sentences, an L2-norm is recommended before calculating the similarity:

python

import torch.nn.functional as F


def similarity(embeddings_1, embeddings_2):
    normalized_embeddings_1 = F.normalize(embeddings_1, p=2)
    normalized_embeddings_2 = F.normalize(embeddings_2, p=2)
    return torch.matmul(
        normalized_embeddings_1, normalized_embeddings_2.transpose(0, 1)
    )


print(similarity(english_embeddings, italian_embeddings))
print(similarity(english_embeddings, japanese_embeddings))
print(similarity(italian_embeddings, japanese_embeddings))

Details

Details about data, training, evaluation and performance metrics are available in the original paper.

BibTeX entry and citation info

bibtex

@inproceedings{mao-nakagawa-2023-lealla,
    title = "{LEALLA}: Learning Lightweight Language-agnostic Sentence Embeddings with Knowledge Distillation",
    author = "Mao, Zhuoyuan  and
      Nakagawa, Tetsuji",
    booktitle = "Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.eacl-main.138",
    doi = "10.18653/v1/2023.eacl-main.138",
    pages = "1886--1894",
    abstract = "Large-scale language-agnostic sentence embedding models such as LaBSE (Feng et al., 2022) obtain state-of-the-art performance for parallel sentence alignment. However, these large-scale models can suffer from inference speed and computation overhead. This study systematically explores learning language-agnostic sentence embeddings with lightweight models. We demonstrate that a thin-deep encoder can construct robust low-dimensional sentence embeddings for 109 languages. With our proposed distillation methods, we achieve further improvements by incorporating knowledge from a teacher model. Empirical results on Tatoeba, United Nations, and BUCC show the effectiveness of our lightweight models. We release our lightweight language-agnostic sentence embedding models LEALLA on TensorFlow Hub.",
}